A method and system for predicting the travel demand of electric official vehicles

Abstract

This application relates to the field of electric official vehicle travel demand management technology, and discloses a method and system for predicting electric official vehicle travel demand. The method includes collecting data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, meteorological information, and traffic flow of electric official vehicles to obtain raw data aligned with a unified timestamp; performing missing value imputation, outlier removal, and time series resampling based on the raw data to form a first dataset. By using empirical mode decomposition to decompose multi-source non-stationary time-series signals, employing arithmetic optimization algorithms to dynamically optimize the parameters of the long short-term memory network, and performing multi-period travel and energy consumption prediction based on the optimized long short-term memory network model, the method addresses the problems of unstable prediction accuracy, insufficient parameter adaptability, and poor model convergence in existing technologies under multi-source nonlinear data environments.

Description

Technical Field

[0001] This invention relates to the field of electric official vehicle travel demand management technology, specifically to a method and system for predicting electric official vehicle travel demand. Background Technology

[0002] Existing methods for predicting the travel demand of electric official vehicles mainly rely on statistical regression, machine learning, or deep learning models. Statistical methods cannot effectively capture nonlinear features when processing multi-source time-series data. Although machine learning methods can identify complex patterns, they are prone to overfitting under multi-dimensional inputs. Deep learning models, such as Long Short-Term Memory networks, have strong time-series learning capabilities, but are still limited by fixed parameters and noise interference from input signals, resulting in large fluctuations in prediction accuracy and insufficient model stability. Especially in the scenario of electric official vehicles, travel demand and energy consumption are coupled with multiple factors such as weather, traffic, and task scheduling. The data is non-stationary and significantly noisy, making it difficult for traditional models to maintain the continuity and interpretability of prediction results.

[0003] Therefore, the problem to be solved by this invention is: how to realize a calculation method that can adaptively decompose signal features, dynamically optimize model parameters, and stably output future travel demand and energy consumption prediction results under the condition of multi-source, non-stationary electric vehicle operation data. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for predicting the travel demand of electric official vehicles, which solves the problems of unstable prediction accuracy, insufficient parameter adaptation capability, and poor model convergence in multi-source nonlinear data environments.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting the travel demand of electric official vehicles, comprising the following steps:

[0006] S1. By collecting data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, weather information, and traffic flow of electric official vehicles, raw data aligned with a unified timestamp is obtained; the first dataset is formed based on the raw data.

[0007] S2. By performing correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence, feature variables with correlation coefficients not lower than a preset lower threshold are identified; variance inflation calculations are performed on the identified feature variables, and collinear variables are removed according to the preset threshold to form a second feature set;

[0008] S3. Perform empirical mode decomposition on the time series in the second feature set to obtain the intrinsic mode function sequence and residual sequence; perform energy proportion statistics and sample entropy statistics on each intrinsic mode function, and select the retained intrinsic mode functions according to the lower and upper limits of the energy proportion interval and the lower and upper limits of the sample entropy interval; sum the retained intrinsic mode functions according to their weights to generate the third time series matrix.

[0009] S4. By constructing the structural parameters and training parameters of the Long Short-Term Memory network into a parameter vector, a parameter space is formed that includes the number of hidden units, the number of network layers, the learning rate, the sliding window length, and the target weight coefficients. Using the iterative update rules preset by the arithmetic optimization algorithm, global optimization is performed on the parameter vector, and fitness calculation and parameter selection are performed with the root mean square error of the validation set as the objective function. The optimal parameter vector is solved and the optimized parameter set is output.

[0010] S5. By inputting the third time series matrix and the optimized parameter set into the long short-term memory network, forward computation and training update are performed to solve the mileage sequence and power consumption sequence within the future predetermined time. The solved prediction sequence is output according to the time index to generate the prediction result of electric official vehicle travel demand.

[0011] Preferably, in S1, the driving mileage refers to the driving distance calculated by the vehicle odometer or GPS track between adjacent timestamps, in kilometers, forming a driving mileage sequence aligned by timestamps;

[0012] The travel time refers to an interval-type record consisting of the departure timestamp and arrival timestamp of each task;

[0013] The charging record includes a start timestamp, an end timestamp, charging amount in kilowatt-hours, initial state of charge percentage, end state of charge percentage, charging station identifier, charging power in kilowatts, and session identifier, all aligned by timestamps.

[0014] Battery capacity refers to the vehicle's rated capacity and usable capacity, measured in kilowatt-hours.

[0015] Unit energy efficiency refers to the energy intensity index of a vehicle, including the kilowatt-hours of electricity consumed per 100 kilometers, and is calculated as kilowatt-hours per kilometer according to the conversion relationship between energy consumption and driving distance.

[0016] Meteorological information includes temperature, relative humidity, precipitation, wind speed, wind direction, solar irradiance, and air pressure, and is spatially mapped or matched to the nearest location based on vehicle or station location to ensure that its time index is consistent with vehicle operation data;

[0017] Traffic flow refers to a sequence of traffic states constructed based on road location and time index. Each time step in the traffic state sequence includes the number of vehicles passing through the road, the average speed, and traffic event markers.

[0018] Preferably, S2 includes:

[0019] By matching all time series variables in the first dataset with the mileage and power consumption sequences in the target sequence according to a unified time index, each variable in mileage, travel time, charging records, battery capacity, unit energy efficiency, meteorological information and traffic flow is used as a feature variable to form a feature correlation calculation set containing a one-to-one correspondence between each feature variable and the target sequence.

[0020] The power consumption sequence includes the actual power consumption of the vehicle's charging records at each time step, arranged with a unified time index. The value is determined by subtracting the feedback power consumption in the same time period from the increase in the charging power in the vehicle's charging records at each time step, and then comparing and verifying the driving mileage with the unit energy efficiency.

[0021] By calculating the Pearson correlation coefficient and partial correlation coefficient for each feature variable in the feature correlation calculation set and the target sequence respectively, correlation calculations are performed under linear and control conditions.

[0022] The Pearson correlation coefficient is calculated by calculating the mean, standard deviation, and covariance of the feature variables and the target sequence at all time steps, and using the ratio of the product of the covariance and the standard deviation as the correlation coefficient. The calculated correlation coefficient is compared with a preset lower threshold, and feature variables with correlation coefficients not lower than the preset lower threshold are selected to generate an initial feature set.

[0023] Preferably, step S2 further includes:

[0024] Variance inflation is calculated for each feature variable in the initial feature set. A linear regression equation is established by using this feature variable as the dependent variable and the remaining feature variables as independent variables, and the coefficient of determination is calculated. and according to the formula Calculate the variance inflation value;

[0025] Feature variables with variance inflation values ​​exceeding a preset threshold are identified as collinear variables, which are used to identify features with multiple linear dependencies in the explanatory space.

[0026] Features identified as collinear variables are removed from the initial feature set, while features with variance inflation values ​​below a threshold are retained.

[0027] The retained feature variables are re-aligned and standardized according to the time index to form a second feature set.

[0028] Preferably, step S3 includes:

[0029] By arranging each time series in the second feature set according to a uniform time index to form a continuous input signal sequence, and performing local extremum identification at each time step, the distribution boundaries of the upper and lower envelopes of the signal are determined.

[0030] The upper and lower envelopes are constructed as continuous functions through spline interpolation, and their mean functions at each time step are calculated.

[0031] The local fluctuation signal is obtained by subtracting the mean function from the original input signal, and the extreme value detection and mean calculation are repeated until the change of the mean function of the local fluctuation signal is lower than the preset convergence threshold, thereby solving the first eigenmode function.

[0032] After separating the first intrinsic mode function from the input signal, its residual part is retained as a new input signal, and the sieving and iteration process continues until the residual signal no longer contains local extrema, thereby solving the entire sequence of intrinsic mode functions and the final residual sequence in sequence.

[0033] Preferably, step S3 further includes:

[0034] For each intrinsic mode function, calculate its energy percentage and sample entropy. The energy percentage is calculated as the ratio of the sum of squares of the mode function to the sum of squares of all mode functions. The sample entropy is calculated as the average negative logarithm of the self-similar probability logarithms based on time delay embedding.

[0035] Eigenmode functions whose energy percentage is between the lower and upper limits of a preset interval and whose sample entropy is between the lower and upper limits of a preset steady-state interval are marked as reserved modes.

[0036] The eigenmode functions marked as preserved modes are weighted and summed using their energy percentages as weighting coefficients to generate a third time series matrix, which is then output.

[0037] Preferably, step S4 includes:

[0038] By defining the structural parameters and training parameters of the Long Short-Term Memory network as adjustable variables, the structural parameters include the number of hidden units, the number of network layers, and the sliding window length, while the training parameters include the learning rate and the target weight coefficient.

[0039] All parameter variables are encoded using floating-point numbers within their allowed value range to form a parameter vector, and the set of parameter vectors is defined as the parameter space.

[0040] A fixed number of candidate parameter vectors are randomly generated in the parameter space, and each parameter vector corresponds to a search individual.

[0041] For each search entity, set position and direction variables to represent the entity's current value and update direction in the parameter space, respectively.

[0042] Store the initial search population as the first generation population and initialize the arithmetic operation control coefficients.

[0043] Preferably, step S4 further includes:

[0044] In each iteration, the globally optimal parameter vector in the current population is selected as a reference, and its value is successively substituted into the arithmetic update rules of addition, subtraction, multiplication and division operations.

[0045] The addition and subtraction operations are used to adjust the precision of parameters within a local range, while the multiplication and division operations are used to expand the search area globally.

[0046] By performing a combination of addition, subtraction, multiplication, and division updates, each candidate parameter vector generates a new position vector in the current iteration step, forming a new group of candidate parameters;

[0047] Each updated candidate parameter vector is input into the long short-term memory network model, the network structure is fixed and forward propagation is performed, and the root mean square error on the validation set is calculated.

[0048] Calculate the fitness value for each parameter vector, where fitness is defined as the reciprocal of the root mean square error;

[0049] Sort all parameter vectors from high to low fitness values, retain parameter vectors with fitness values ​​greater than a preset threshold as the next generation of candidate parameters, and remove parameter vectors with fitness values ​​less than the preset threshold from the group.

[0050] Preferably, step S4 further includes:

[0051] The arithmetic optimization algorithm iteration stops when the number of iterations reaches the preset upper limit or when the change in the global optimal fitness is lower than the preset convergence threshold in multiple consecutive iterations.

[0052] Define the parameter vector with the highest current fitness as the optimal parameter vector, and solve for the value of each parameter variable in it to form an optimized parameter set;

[0053] The optimized parameter set is output and written into the structure and training configuration of the Long Short-Term Memory network, serving as the input parameters for subsequent model training and prediction calculations.

[0054] A system for predicting the travel demand of electric official vehicles, comprising:

[0055] The data construction module collects data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, weather information, and traffic flow of electric official vehicles to obtain raw data aligned with a unified timestamp; and forms the first dataset based on the raw data.

[0056] The feature selection module performs correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence to identify feature variables with correlation coefficients not lower than a preset lower threshold; it then performs variance inflation calculations on the identified feature variables and removes collinear variables according to a preset threshold to form a second feature set.

[0057] The sequence decomposition and reconstruction module performs empirical mode decomposition on the time series in the second feature set to solve for the intrinsic mode function sequence and the residual sequence; it performs energy proportion statistics and sample entropy statistics on each intrinsic mode function, and selects the retained intrinsic mode functions based on the lower and upper limits of the energy proportion interval and the lower and upper limits of the sample entropy interval; it then sums the retained intrinsic mode functions according to their weights to generate the third time series matrix.

[0058] The parameter optimization module constructs a parameter vector from the structural parameters and training parameters of the Long Short-Term Memory network, forming a parameter space containing the number of hidden units, the number of network layers, the learning rate, the sliding window length, and the target weight coefficients. It then performs global optimization on the parameter vector using the iterative update rules preset by the arithmetic optimization algorithm, and calculates the fitness and selects parameters with the root mean square error of the validation set as the objective function. Finally, it solves for the optimal parameter vector and outputs the optimized parameter set.

[0059] The prediction generation module inputs the third time series matrix and the optimized parameter set into the long short-term memory network, performs forward computation and training updates, and solves the mileage sequence and power consumption sequence within a predetermined future time period; the solved prediction sequence is output according to the time index to generate the prediction result of electric official vehicle travel demand.

[0060] This invention provides a method and system for predicting the travel demand of electric official vehicles. It has the following beneficial effects:

[0061] 1. This invention effectively solves the problems of large fluctuations in prediction accuracy and insufficient adaptive ability caused by fixed parameters when the traditional electric official vehicle travel demand prediction method is used to process multi-source non-stationary data by introducing a collaborative mechanism of empirical mode decomposition and arithmetic optimization algorithm, thereby achieving stable prediction output in complex coupled environments.

[0062] 2. This invention systematically eliminates redundant and collinear variables by performing correlation calculation and variance inflation analysis during the feature construction stage, ensuring the independence and identifiability of input features, reducing multiple correlation interferences during model training, and improving the statistical robustness of feature input.

[0063] 3. This invention performs adaptive signal decomposition on time series through empirical mode decomposition, decomposing the original complex signal into multi-scale intrinsic mode functions and selecting and retaining modes, thereby achieving hierarchical analysis of local energy characteristics and temporal complexity in non-stationary sequences, thus improving the model's ability to capture dynamic trends.

[0064] 4. This invention uses an arithmetic optimization algorithm to globally optimize the parameters of the Long Short-Term Memory network. By utilizing arithmetic update rules of addition, subtraction, multiplication, and division and adaptive adjustment of dynamic control coefficients, it achieves joint optimization of structural parameters and training parameters, enabling the model to maintain a balance between global search and local convergence, thereby improving the prediction accuracy and convergence stability of the network.

[0065] 5. This invention continuously advances along the time axis through a sliding window prediction mechanism, dynamically splicing the predicted output to the end of the input sequence, thereby realizing continuous prediction output of travel mileage and energy consumption for multiple time periods in the future. This provides a quantifiable decision-making basis for the scheduling management, energy consumption planning and charging strategy of electric official vehicles. Attached Figure Description

[0066] Figure 1 This is a flowchart of the method of the present invention.

[0067] Figure 2 This is a system architecture diagram of the present invention. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Example:

[0070] Please see the appendix Figure 1 - Appendix Figure 2 This invention provides a method for predicting the travel demand of electric official vehicles, comprising the following steps:

[0071] S1. Data Construction:

[0072] By collecting data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, weather information, and traffic flow of electric official vehicles, raw data aligned with a unified timestamp is obtained; based on the raw data, missing value imputation, outlier removal, and time series resampling are performed to form the first dataset;

[0073] Additionally, it should be noted that missing value imputation refers to generating complete values ​​by interpolation or averaging of adjacent time periods when there are missing time steps or numerical gaps in the original data, in order to maintain the continuity of the time series.

[0074] Outlier removal refers to identifying values ​​that exceed the normal statistical range (including but not limited to points exceeding three standard deviations of the mean) and replacing them with adjacent valid values ​​or reestimating them based on the median to eliminate abnormal fluctuations.

[0075] Time series resampling refers to remapping data with different sampling frequencies or irregular time intervals to a uniform time step (such as every 15 minutes or every hour), and calculating the mean or sum within each new time step to form an evenly spaced time series;

[0076] S2. Feature Filtering: By performing correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence, feature variables with correlation coefficients not lower than a preset lower threshold are identified; variance inflation calculations are performed on the identified feature variables, and collinear variables are removed according to a preset threshold to form a second feature set;

[0077] S3. Sequence Decomposition and Reconstruction:

[0078] By performing empirical mode decomposition on the time series in the second feature set, the intrinsic mode function sequence and the residual sequence are obtained;

[0079] Perform energy percentage statistics and sample entropy statistics on each intrinsic mode function, and select and retain intrinsic mode functions based on the lower and upper limits of the energy percentage interval and the lower and upper limits of the sample entropy interval.

[0080] The retained intrinsic mode functions are summed by weights to generate the third time series matrix;

[0081] S4. Parameter optimization:

[0082] By constructing the structural parameters and training parameters of the Long Short-Term Memory network into a parameter vector, a parameter space is formed that includes the number of hidden units, the number of network layers, the learning rate, the sliding window length, and the target weight coefficients.

[0083] Using the iterative update rules of addition, subtraction, multiplication and division operations preset by the arithmetic optimization algorithm, global optimization is performed on the parameter vector, and fitness calculation and parameter selection are performed with the root mean square error of the validation set as the objective function.

[0084] Solve for the optimal parameter vector and output the optimized parameter set for configuring the prediction model.

[0085] S5, Prediction Generation:

[0086] By inputting the third time series matrix and the optimized parameter set into the long short-term memory network, forward computation and training updates are performed to solve the driving mileage sequence and power consumption sequence within a predetermined future time period;

[0087] The solved prediction sequences are output according to time index to generate prediction results for the travel demand of electric official vehicles.

[0088] For S5, it should be noted that in the prediction generation process, the system first inputs the third time sequence matrix as the input signal sequence into the input layer of the long short-term memory network, and at the same time calls the number of hidden units, the number of network layers, the learning rate, the sliding window length and the target weight coefficient in the optimization parameter set to configure the network structure and training process.

[0089] During the forward computation phase, the system feeds the input sequence into the Long Short-Term Memory (LSTM) network unit sequentially, indexed by time step, and performs state updates for the input gate, forget gate, and output gate in turn.

[0090] The input gate controls the proportion of the current input information written, the forget gate determines the proportion of the previous state information to be retained, and the output gate adjusts the weight of the current state in the output.

[0091] The system calculates the cell state and hidden state at each time step and propagates them throughout the time series in a recursive manner, finally generating the predicted value for the corresponding time step at the output layer.

[0092] During the training and update phase, the system calculates the mean squared error and performs backpropagation based on the error between the predicted output and the real target sequence (mileage sequence and power consumption sequence). It adjusts the gradient descent step size by optimizing the learning rate in the parameter set and updates the weight matrix and bias parameters layer by layer until the training loss function decreases below the preset threshold or reaches the maximum training rounds, thus obtaining a converged long short-term memory network model.

[0093] During the prediction phase, the system fixes the converged network parameters and uses the end window of the third time series matrix as the initial input to perform sliding window prediction:

[0094] For each time step prediction output generated, it is concatenated to the end of the input sequence and the earliest time step input is removed. Then, the prediction of the next time step is continued, thereby gradually solving the driving mileage sequence and power consumption sequence within the future predetermined time step.

[0095] The system aligns and reconstructs the two sets of prediction sequences according to a unified time index to form the prediction results of electric official vehicle travel demand.

[0096] The results include cumulative driving demand, energy consumption trends, and a multi-period prediction index matrix for future time intervals, which can be used to support subsequent scheduling, charging, and energy consumption optimization decisions.

[0097] Secondly, in this scheme, the electric official vehicle travel demand forecast result refers to the set of quantitative results generated by the system after completing the Long Short-Term Memory Network forecast calculation, which describes the state of vehicle travel and energy consumption demand in the future period. This result specifically consists of two types of core forecast sequences and their derived indicators:

[0098] First, the mileage sequence for each time period in the future, representing the cumulative mileage requirement for all official vehicles within a continuous future time period;

[0099] Second, the power consumption sequence corresponding to the time step represents the total power consumption required to meet the travel demand. After generating the above two time series, the system further aligns and calculates them according to a unified time index to form a comprehensive travel demand prediction matrix. This matrix includes fields such as travel distance, power consumption, average energy intensity, and demand growth rate for each time step. After summarizing and calculating, the system can output the cumulative travel distance, cumulative power consumption, and energy consumption change trend within a specified future time interval. This is used to intuitively reflect the travel intensity and energy demand level of the future official vehicle fleet, thereby providing data basis for scheduling planning, charging pile allocation, and energy consumption optimization.

[0100] In S1, the driving mileage refers to the distance traveled between adjacent timestamps, calculated by the vehicle odometer or GPS trajectory, in kilometers, forming a driving mileage sequence aligned by timestamps;

[0101] The travel time refers to an interval-type record consisting of the departure timestamp and arrival timestamp of each task, and a binary travel state sequence is generated on equally spaced time steps for alignment.

[0102] The methods for generating binary execution state sequences include:

[0103] The departure and arrival timestamps of each electric official vehicle are mapped to a globally unified timeline. If the time step is within any travel interval within each fixed time interval, it is assigned a value of 1; otherwise, it is assigned a value of 0, thus forming a continuous binary execution state sequence.

[0104] Charging records refer to records based on charging sessions, including start timestamp, end timestamp, charging amount in kilowatt-hours, initial state of charge percentage, end state of charge percentage, charging station identifier, charging power in kilowatts, and session identifier, all aligned by timestamps.

[0105] The record based on a charging session refers to a complete vehicle charging process as the recording unit, that is, the continuous charging process from the moment the vehicle is connected to the charging pile to the moment the charging is completed and the connection is disconnected.

[0106] A session identifier is a unique number used to identify each charging session, distinguishing independent charging records generated at different times, by different vehicles, or by different charging stations.

[0107] Battery capacity refers to the vehicle's rated capacity and available capacity, measured in kilowatt-hours, and is used as a vehicle-level static parameter for energy consumption calculation and constraints.

[0108] Unit energy efficiency refers to the energy intensity index of a vehicle, including the electricity consumption per 100 kilometers as a measure, and calculated as per kilometer according to the conversion relationship between energy consumption and driving distance, which is used to quantify driving mileage and electricity consumption in a unified manner.

[0109] It should be noted that the method for calculating the kilometer per kilowatt-hour based on the conversion relationship between energy consumption and driving distance includes: dividing the total driving distance of the vehicle within a given time period by the total power consumption within that time period to calculate the distance that can be traveled per unit of electricity, i.e., kilometer per kilowatt-hour; meteorological information includes temperature, relative humidity, precipitation, wind speed, wind direction, solar irradiance, and air pressure, and spatial mapping or nearest matching is performed according to the vehicle location or station location to ensure that its time index is consistent with the vehicle operation data;

[0110] The implementation methods for spatial mapping or proximity matching based on vehicle or station location include: calculating the distance between the vehicle positioning coordinates or charging station coordinates and the meteorological monitoring grid or meteorological station coordinates, and selecting the data of the nearest meteorological monitoring point as the meteorological information corresponding to the vehicle or station at each time step; traffic flow refers to the traffic state sequence constructed based on road location and time index, and each time step in the traffic state sequence includes the number of vehicles passing through the road, the average driving speed, and traffic event markers, which are used to characterize the road traffic conditions.

[0111] S2 includes:

[0112] By matching all time-series variables in the first dataset with the mileage and power consumption sequences in the target sequence using a unified time index, each variable among mileage, travel time, charging records, battery capacity, unit energy efficiency, meteorological information, and traffic flow is used as a feature variable, forming a feature correlation calculation set containing a one-to-one correspondence between each feature variable and the target sequence. The power consumption sequence includes the actual power consumption of the vehicle's charging records at each time step, arranged with a unified time index. The value is determined by subtracting the feedback power consumption in the same time period from the increase in the vehicle's charging records in each time step, and then comparing and verifying with mileage and unit energy efficiency. The unit is kilowatt-hour. In addition, when determining the value after comparing and verifying with mileage and unit energy efficiency, the power consumption calculated from the charging records can be compared with the theoretical power consumption obtained by multiplying mileage by unit energy efficiency in the same time period. When the difference between the two is within a set error range, the average value is taken as the final power consumption of that time step; otherwise, the actual power consumption of the charging records is used as the corrected value.

[0113] All time series variables in the first dataset refer to various variable sequences that remain time-continuous and arranged according to a unified time step index after missing value imputation, outlier removal, and time series resampling. These include mileage sequences, travel time sequences, charging record sequences, battery capacity sequences, unit energy efficiency sequences, meteorological information sequences, and traffic flow sequences. These variables are all indexed by time steps, forming a complete set of time series that can be used for correlation calculation. The target sequence refers to the time series data used to predict the output target of the model. In this scheme, it includes the mileage sequence and power consumption sequence representing the vehicle operation results.

[0114] The correlation calculation is performed under linear and control conditions by calculating the Pearson correlation coefficient and partial correlation coefficient for each feature variable and the target sequence in the feature correlation calculation set. The Pearson correlation coefficient is calculated by calculating the mean, standard deviation and covariance of the feature variable and the target sequence at all time steps, and using the ratio of the product of covariance and standard deviation as the correlation coefficient to obtain the correlation value reflecting the strength of the linear correlation between variables. The calculated correlation coefficient is compared with a preset lower threshold, and feature variables with correlation coefficients not lower than the preset lower threshold are selected to generate an initial feature set.

[0115] Variance inflation is calculated for each feature variable in the initial feature set. A linear regression equation is established by using this feature variable as the dependent variable and the remaining feature variables as independent variables, and the coefficient of determination is calculated. and according to the formula Calculate the variance inflation value; identify feature variables with variance inflation values ​​higher than a preset threshold as collinear variables, which are used to identify features with multiple linear dependencies in the interpretation space;

[0116] The principle behind variance inflation calculations lies in quantitatively analyzing the linear dependence between different characteristic variables to determine whether there is redundancy or collinearity among the characteristics. Specifically, a linear regression equation is established using one characteristic variable as the dependent variable and the remaining characteristic variables as independent variables to test whether the dependent variable can be explained by a linear combination of the other characteristics. If the coefficient of determination of the regression equation is... A high variance indicates that the change in this feature can be well fitted by other features, thus indicating multiple linear dependence. To quantify the degree of dependence, the variance inflation value is calculated, which is done as follows: When the variance inflation value exceeds the preset threshold, it indicates that the feature is highly correlated with other features in the explanatory space. Continuing to retain it will cause information redundancy and interfere with model learning. Therefore, it is identified as a collinear variable and removed. The purpose of this calculation process is to ensure that the input features are independent of each other, reduce the interference of multiple correlations on the training of model weights, and thus improve the stability and generalization ability of the prediction model.

[0117] Features identified as collinear variables are removed from the initial feature set, while features with variance inflation values ​​below a threshold are retained. The retained features are then re-aligned and standardized according to their time indices to form a second feature set.

[0118] The reason for this approach is that removing feature variables with variance inflation values ​​exceeding a threshold is to eliminate feature redundancy caused by multiple linear dependencies in the dataset. In practice, the system directly deletes features exceeding the threshold from the initial feature set based on the variance inflation value of each feature, retaining only features with variance inflation values ​​below the threshold. These retained features are then rearranged according to a uniform time index order. Finally, a standardization transformation with a mean of zero and a variance of one is performed on the values ​​of each feature, ensuring that all features participate in subsequent calculations on the same numerical scale. The purpose of this process is to ensure that the input features are numerically comparable and statistically independent.

[0119] S3 includes:

[0120] By arranging each time series in the second feature set according to a unified time index, a continuous input signal sequence is formed. Local extremum identification is performed at each time step to determine the distribution boundaries of the upper and lower envelopes of the signal. Each time series in the second feature set refers to the continuous value sequence of the feature variable in the time dimension after being filtered by correlation calculation and variance inflation calculation. Each sequence is arranged with a unified time index to represent the numerical change of the feature at each time step. The input signal sequence refers to the continuous signal set formed by arranging the time series in the second feature set according to the time index and then aggregating the values ​​of each feature at the same time step. This signal is used as the input data for empirical mode decomposition to perform local extremum identification and envelope construction.

[0121] Local extremum identification involves iterating through time steps in the input signal sequence, determining the relationship between the value at each time step and the values ​​at the adjacent time steps, marking the current value as a local maximum when it is greater than the values ​​on both sides, and marking the current value as a local minimum when it is less than the values ​​on both sides, thereby determining the set of local extremum points of the signal.

[0122] The upper and lower envelopes are constructed as continuous functions through spline interpolation, and their mean functions at each time step are calculated. The mean function is subtracted from the original input signal to obtain the local fluctuation signal, and the extreme value detection and mean calculation are repeated until the change in the mean function of the local fluctuation signal is lower than the preset convergence threshold, thereby solving the first intrinsic mode function.

[0123] Spline interpolation refers to using the time index of local extreme points as nodes and the corresponding signal values ​​as node values, and employing a cubic spline interpolation function to construct smooth curves between adjacent nodes, thereby forming continuous upper and lower envelope functions.

[0124] The mean function is the sequence of average values ​​of the upper and lower envelope functions at each time step, used to represent the local central tendency of the signal at that time step; the original input signal is the value of the unfiltered input signal sequence at each time step; the local fluctuation signal is the remaining signal obtained by subtracting the mean function from the original input signal, representing the local oscillation component after removing the central tendency; extreme value detection is to re-perform extreme value identification on the local fluctuation signal to update the envelope, and mean calculation is to recalculate the mean function on the updated envelope; by continuously repeating the iterative process of extreme value detection and mean calculation until the change in the mean function is lower than the set convergence threshold, the current local fluctuation signal is the first intrinsic mode function;

[0125] After separating the first intrinsic mode function from the input signal, its residual part is retained as a new input signal, and the sieving and iteration are continued until the residual signal no longer contains local extrema, thereby solving the entire sequence of intrinsic mode functions and the final residual sequence in sequence;

[0126] After completing the first eigenmode function (EMF) sieving iteration, the EMF is separated from the input signal. This is achieved by subtracting the corresponding EMF value from the input signal value at each time step, resulting in a residual signal sequence. The residual signal represents the remaining trend component of the signal after removing the current oscillation component. This residual signal is defined as a new input signal for further decomposition. The system then redefines the sieving iteration operations—local extremum identification, spline interpolation, mean function calculation, and local fluctuation signal update—using the new input signal. In each iteration, a new EMF is solved and the residual signal is updated. When no local extrema exist in the current residual signal and the signal exhibits monotonical variation, the iteration terminates. At this point, the entire sequence of EMFs solved constitutes the multi-scale oscillation component of the signal, while the last residual signal constitutes the final trend component, forming a complete empirical mode decomposition result.

[0127] S3 also includes:

[0128] For each intrinsic mode function, calculate its energy percentage and sample entropy. The energy percentage is calculated as the ratio of the sum of squares of the intrinsic mode function to the sum of squares of all modes. The sample entropy is calculated as the average negative logarithm of the self-similar probability logarithms based on time delay embedding. Intrinsic mode functions whose energy percentage is between the lower and upper limits of a preset interval and whose sample entropy is between the lower and upper limits of a preset steady-state interval are marked as reserved modes.

[0129] For each solved intrinsic mode function (EMF), the system first performs an energy percentage calculation on its complete time series. The sum of the squares of the EMF values ​​at all time steps is taken as the energy value of the EMF. This energy value is then compared with the sum of the energy values ​​of all EMFs to obtain the energy percentage of the EMF in the overall signal. The energy percentage characterizes the contribution of the EMF to the overall signal energy. Subsequently, the system performs sample entropy calculation on the same EMF. Sample entropy is defined as the average negative logarithm of the self-similarity probability based on time delay embedding, used to quantify the EMF. The system addresses the time complexity of the intrinsic mode functions (IMFs). Specifically, it constructs vector subsequences of IMFs over the time series based on the embedding dimension and time delay. It calculates the matching probability when the distance between adjacent subsequences is less than the tolerance threshold, and takes the negative of the logarithmic mean of all matching probabilities to obtain the sample entropy value. After completing the above calculations, the system performs interval filtering on the energy percentage and sample entropy value of each IMF. Only when the energy percentage is between the lower and upper limits of the preset energy interval and the sample entropy is between the lower and upper limits of the preset steady-state interval is the IMF marked as a reserved mode for subsequent weighted reconstruction.

[0130] The eigenmode functions marked as preserved modes are weighted and summed using their energy proportions as weighting coefficients to generate a third time series matrix, which is then output as multi-scale input data for subsequent modeling.

[0131] For intrinsic mode functions (IMFs) already marked as preserved modes, a set of weighted coefficients is first constructed using the energy proportion of each IMF. This set of weighted coefficients is then normalized by the sum of the energy proportions of the preserved modes, making each weighted coefficient numerically independent of time. Subsequently, at each time index, the corresponding time point values ​​of the preserved IMFs are multiplied by the normalized weighted coefficients and summed to obtain the reconstructed values ​​for the corresponding time points, forming a reconstructed sequence arranged by time index. The above weighted reconstruction is then performed on each time series in the second feature set, and the reconstructed sequences are aligned and concatenated column-wise to generate a third time series matrix, which is then output as multi-scale input data for subsequent modeling.

[0132] S4 includes:

[0133] By defining the structural parameters and training parameters of the Long Short-Term Memory network as adjustable variables, the structural parameters include the number of hidden units, the number of network layers, and the sliding window length, while the training parameters include the learning rate and the target weight coefficient; all parameter variables are encoded as floating-point numbers within their allowed value range to form parameter vectors, and the set of parameter vectors is defined as the parameter space for iterative updates of the arithmetic optimization algorithm;

[0134] For the structural and training parameters of the Long Short-Term Memory (LSTM) network, the system first defines them as a continuously adjustable set of variables. The structural parameters control the hierarchical shape of the network topology, including the number of hidden units, the number of network layers, and the sliding window length. The training parameters adjust the update step size and error distribution during the learning process, including the learning rate and the target weight coefficient. The system uses the physical or empirical range of each parameter as a boundary, dividing the value interval between the minimum and maximum values ​​according to a set precision. Each parameter is numerically encoded in floating-point form, forming a real-number vector containing all parameters; each vector corresponds to a complete network configuration scheme. All encoded parameter vectors are combined into a parameter vector set, which is defined as the parameter space. This set is used by the arithmetic optimization algorithm to perform search and update operations during iteration, ensuring that each parameter update occurs within a continuous domain of the parameter space, thereby guaranteeing the computability and continuous convergence of the parameter optimization process.

[0135] A fixed number of candidate parameter vectors are randomly generated in the parameter space, and each parameter vector corresponds to a search individual; position variables and direction variables are set for each search individual, representing the current value and update direction of the individual in the parameter space, respectively; the initial search population is stored as the first generation population, and arithmetic operation control coefficients are initialized for proportional control of subsequent arithmetic operations;

[0136] In the aforementioned parameter space, a random number generator is invoked to generate several sets of random floating-point number sequences within the value range of each parameter, and each sequence corresponds to a complete parameter vector. Each parameter vector represents a combination scheme of network structure and training configuration, which the system defines as a search individual. For each search individual, the system sets position and direction variables. The position variable stores the current value state of the individual in the parameter space, and the direction variable stores the movement direction and change magnitude of the individual in subsequent iterations. The set of all search individuals is defined as the first-generation population, used to construct the initial search baseline for the arithmetic optimization algorithm. Subsequently, the system generates arithmetic operation control coefficients, which are calculated by the algorithm's iteration step ratio and random perturbation term, and the formula is defined as follows: ,in and The initial weighting constants are... For the current iteration step, The maximum number of iterations. for Random numbers within the interval; this control coefficient is used to dynamically adjust the update magnitude in subsequent addition, subtraction, multiplication, and division arithmetic operations, so that the arithmetic optimization algorithm maintains global search capability in the early stage of iteration and gradually focuses on the local optimal region in the convergence stage, thereby achieving stable search and continuous optimization of the parameter vector;

[0137] In each iteration, the globally optimal parameter vector in the current population is selected as a reference, and its value is successively substituted into the arithmetic update rules of addition, subtraction, multiplication, and division operations. Among them, addition and subtraction operations are used to adjust the parameter precision in the local range, and multiplication and division operations are used to expand the search area in the global range. By performing the combined update of addition, subtraction, multiplication, and division, each candidate parameter vector generates a new position vector in the current iteration step, forming a new candidate parameter group.

[0138] In each iteration of the arithmetic optimization algorithm, the system first identifies the parameter vector with the highest fitness value in the current population and defines it as the globally optimal parameter vector; this vector serves as the reference benchmark in this round of updates. For each of the remaining candidate parameter vectors, the system sequentially executes the arithmetic operation update rules, combining the components of the globally optimal parameter vector with the components of the current parameter vector, and updating the parameters through a linear combination of addition, subtraction, multiplication, and division operations. Addition and subtraction operations are used to fine-tune the parameter precision within the local range of parameter values; multiplication and division operations are used to expand or shrink the parameter search span to enhance global exploration capabilities. The update magnitude of each arithmetic operation is determined by the arithmetic operation control coefficient, which is calculated based on the iteration step ratio and the random perturbation term, using the formula: ,in and The initial weighting constants are... For the current iteration step, The maximum number of iterations. for Random numbers within the interval; after each update, the system calculates the new position of the parameter vector and replaces the old position with the new position to generate a new generation of candidate parameter groups; this process is repeated in a loop, so that each parameter vector continuously approaches the optimal solution under the dynamic adjustment of the control coefficient, realizing a collaborative search process of global optimization and local convergence;

[0139] Each updated candidate parameter vector is input into the Long Short-Term Memory network model, the network structure is fixed and forward propagation is performed, and the root mean square error on the validation set is calculated. The fitness value of each parameter vector is calculated, and the fitness value is defined as the reciprocal of the root mean square error. All parameter vectors are sorted from high to low fitness value, and parameter vectors with fitness values ​​greater than a preset threshold are retained as the next generation of candidate parameters. Parameter vectors with fitness values ​​less than the preset threshold are removed from the group to achieve dynamic updating and optimization of the group.

[0140] For each candidate parameter vector updated through arithmetic operations, the system sequentially inputs each parameter value from the vector into the Long Short-Term Memory (LSTM) network, maintaining the network's structural configuration and only replacing the parameter vectors for performance evaluation. After each input, the system performs a complete forward propagation process on the validation set, calculating the squared error between the predicted output and the true value at each time step. The system then averages the squared errors over all time steps and takes the square root to obtain the root mean square error (RMSE) corresponding to the parameter vector. Subsequently, the system uses the reciprocal of the RMS error as the fitness value to characterize the prediction accuracy of the parameter vector. The higher the value, the better the prediction effect of the parameter combination under the current network configuration. After completing the fitness calculation of all candidate parameter vectors, the system sorts all parameter vectors from high to low fitness value, and retains parameter vectors with fitness values ​​greater than the threshold to form the next generation of candidate parameter groups, while removing parameter vectors with fitness values ​​less than the threshold from the group. This process ensures that only high-performing parameter combinations are retained in each iteration, so that the arithmetic optimization algorithm group gradually converges to the optimal solution in continuous iterations, thereby achieving global optimization of the parameters of the long short-term memory network.

[0141] S4 also includes:

[0142] When the number of iterations reaches the preset upper limit or the change in the global optimal fitness is lower than the preset convergence threshold in multiple consecutive iterations, the arithmetic optimization algorithm iteration stops; the parameter vector with the highest current fitness is defined as the optimal parameter vector, and the value of each parameter variable in it is solved to form an optimized parameter set; the optimized parameter set is output and written into the structure configuration and training configuration of the long short-term memory network as the input parameter basis for subsequent model training and prediction calculation.

[0143] For the iterative process of the arithmetic optimization algorithm, after each round of updates, the system records the difference between the current global optimal fitness value and the previous optimal fitness value, and determines whether this difference is lower than a preset convergence threshold in multiple consecutive iterations. If this condition is met continuously or the number of iterations reaches a preset upper limit, the system terminates the iterative process of the arithmetic optimization algorithm. After termination, the system selects the parameter vector with the highest fitness value in the current population and defines this parameter vector as the optimal parameter vector. Subsequently, the system reads the parameter component values ​​in this parameter vector one by one, and solves the number of hidden units, the number of network layers, the sliding window length, the learning rate, and the target weight coefficient in sequence, and forms these parameters into an optimized parameter set. Finally, the system writes the optimized parameter set into the structure configuration and training configuration file of the Long Short-Term Memory network, so that the network can directly call these optimal parameters in subsequent model training and prediction calculations, thereby ensuring that the parameters of the model structure and training process are consistent with the optimization results.

[0144] Please see the appendix Figure 2A predictive system for the travel demand of electric official vehicles, comprising:

[0145] The data construction module collects data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, weather information, and traffic flow of electric official vehicles to obtain raw data aligned with a unified timestamp; and forms the first dataset based on the raw data.

[0146] The feature selection module performs correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence to identify feature variables with correlation coefficients not lower than a preset lower threshold; it then performs variance inflation calculations on the identified feature variables and removes collinear variables according to a preset threshold to form a second feature set.

[0147] The sequence decomposition and reconstruction module performs empirical mode decomposition on the time series in the second feature set to solve for the intrinsic mode function sequence and the residual sequence; it performs energy proportion statistics and sample entropy statistics on each intrinsic mode function, and selects the retained intrinsic mode functions based on the lower and upper limits of the energy proportion interval and the lower and upper limits of the sample entropy interval; it then sums the retained intrinsic mode functions according to their weights to generate the third time series matrix.

[0148] The parameter optimization module constructs a parameter vector from the structural parameters and training parameters of the Long Short-Term Memory network, forming a parameter space containing the number of hidden units, the number of network layers, the learning rate, the sliding window length, and the target weight coefficients. It then performs global optimization on the parameter vector using the iterative update rules preset by the arithmetic optimization algorithm, and calculates the fitness and selects parameters with the root mean square error of the validation set as the objective function. Finally, it solves for the optimal parameter vector and outputs the optimized parameter set.

[0149] The prediction generation module inputs the third time series matrix and the optimized parameter set into the long short-term memory network, performs forward computation and training updates, and solves the mileage sequence and power consumption sequence within a predetermined future time period; the solved prediction sequence is output according to the time index to generate the prediction result of electric official vehicle travel demand.

[0150] The working principle of this scheme is executed in four consecutive stages: time series processing, signal decomposition, parameter optimization, and model prediction. First, in the data construction stage, the system collects mileage, travel time, charging records, battery capacity, unit energy efficiency, meteorological information, and traffic flow. After unifying these data to the same time index, missing values ​​are imputed by interpolation or the mean of adjacent time periods. Outliers are identified and replaced according to a three-standard-deviation rule. The system is then resampled at a fixed step size to form the first dataset. Subsequently, in the feature selection stage, the first dataset is aligned with the mileage and energy consumption sequences in the target sequence. Pearson correlation coefficients and partial correlation coefficients are calculated for each variable. Variables with correlation coefficients not lower than a threshold are retained, and variance inflation is calculated for each selected variable. Collinear variables with variance inflation values ​​exceeding a threshold are removed to obtain the second feature set. In the sequence decomposition and reconstruction stage, the system performs empirical mode decomposition on each time series in the second feature set, obtains the upper and lower envelopes through extreme point spline interpolation and iteratively filters them, and solves the intrinsic mode function sequence and residual sequence in turn. The energy proportion and sample entropy of each intrinsic mode function are calculated, and the retained functions are selected according to the preset interval. The energy proportion is normalized as the weight and weighted summation is performed at each time step to generate the third time series matrix.

[0151] In the parameter optimization stage, the system constructs a parameter space by combining the number of hidden units, network layers, learning rate, sliding window length, and target weight coefficients of the Long Short-Term Memory (LSTM) network. It then uses arithmetic optimization algorithms with addition, subtraction, multiplication, and division update rules and control coefficients for a group iterative search. The root mean square error of the validation set is used as the fitness criterion, and the optimized parameter set is retained at a threshold and output when convergence or the maximum number of steps is reached. Finally, in the prediction generation stage, the third time series matrix and the optimized parameter set are input into the LSM network. It is first trained to early stopping using the target weighted mean square error, and then the mileage and power consumption sequences for each future time step are solved sequentially using a sliding window recursive method. The results are then output according to a unified time index to form the prediction results for electric official vehicle travel demand.

Claims

1. A method for predicting the travel demand of electric official vehicles, characterized in that, Includes the following steps: S1. By collecting data on electric official vehicles, including vehicle mileage data, charging data, and weather and traffic data of the vehicle's location, raw data aligned with a unified timestamp is obtained; the first dataset is formed based on the raw data. S2. By performing correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence, feature variables with correlation coefficients not lower than a preset lower threshold are identified; variance inflation calculations are performed on the identified feature variables, and collinear variables are removed according to the preset threshold to form a second feature set; S3. By performing empirical mode decomposition on the time series in the second feature set, the intrinsic mode function sequence and residual sequence are obtained; energy proportion statistics and sample entropy statistics are performed on each intrinsic mode function, and the intrinsic mode functions to be retained are selected according to the lower and upper limits of the energy proportion interval and the lower and upper limits of the sample entropy interval. The retained intrinsic mode functions are summed by weights to generate the third time series matrix; S4. Construct a parameter vector by combining the structure and training hyperparameters of the long short-term memory network. The parameter vector is globally optimized using an arithmetic optimization algorithm, with the prediction error of the model on the validation set as the objective function. The optimal parameter vector is solved, and the optimized parameter set is output. The structural parameters and training parameters of the Long Short-Term Memory (LSTM) network are defined as adjustable variables. Structural parameters include the number of hidden units, the number of network layers, and the sliding window length. Training parameters include the learning rate and the target weight coefficient. All parameter variables are encoded as floating-point numbers within their allowed value range to form parameter vectors, and the set of parameter vectors is defined as the parameter space. A fixed number of candidate parameter vectors are randomly generated within the parameter space, with each parameter vector corresponding to a search individual. Position and direction variables are set for each search individual, representing the individual's current value and update direction in the parameter space, respectively. The initial search population is stored as the first-generation population, and arithmetic operation control coefficients are initialized. In each iteration, the globally optimal parameter vector in the current population is selected as a reference, and its value is sequentially substituted into the arithmetic update rules of addition, subtraction, multiplication, and division operations. Addition and subtraction operations are used to adjust parameter precision locally, while multiplication... The arithmetic and division operations are used to expand the search area globally. By performing a combination of addition, subtraction, multiplication, and division to update, each candidate parameter vector generates a new position vector in the current iteration step, forming a new candidate parameter group. Each updated candidate parameter vector is input into the Long Short-Term Memory network model, the network structure is fixed, and forward propagation is performed. The root mean square error on the validation set is calculated, and the fitness value of each parameter vector is calculated. Fitness is defined as the reciprocal of the root mean square error. All parameter vectors are sorted from high to low fitness values. Parameter vectors with fitness values ​​greater than a preset threshold are retained as the next generation of candidate parameters, and parameter vectors with fitness values ​​less than the preset threshold are removed from the group. When the number of iterations reaches a preset upper limit or the change in the global optimal fitness is lower than the preset convergence threshold in multiple consecutive iterations, the arithmetic optimization algorithm iteration is stopped. The parameter vector with the highest current fitness is defined as the optimal parameter vector, and the value of each parameter variable in it is solved to form an optimized parameter set. The optimized parameter set is output and written into the structure configuration and training configuration of the Long Short-Term Memory network as the input parameters for subsequent model training and prediction calculations. S5. By inputting the third time series matrix and the optimized parameter set into the long short-term memory network, forward computation and training update are performed to solve the mileage sequence and power consumption sequence within the future predetermined time. The solved prediction sequence is output according to the time index to generate the prediction result of electric official vehicle travel demand.

2. The method for predicting the travel demand of electric official vehicles according to claim 1, characterized in that, In step S1, the driving mileage refers to the driving distance calculated by the vehicle odometer or GPS track between adjacent timestamps, in kilometers, forming a driving mileage sequence aligned by timestamps. The travel time refers to the interval record formed by the departure timestamp and arrival timestamp of each task. The charging record includes a start timestamp, an end timestamp, charging amount in kilowatt-hours, initial state of charge percentage, end state of charge percentage, charging station identifier, charging power in kilowatts, and session identifier, all aligned by timestamps. Battery capacity refers to the vehicle's rated capacity and usable capacity, measured in kilowatt-hours. Unit energy efficiency refers to the energy intensity index of a vehicle, including the kilowatt-hours of electricity consumption per 100 kilometers, and is calculated as kilowatt-hours per kilometer according to the conversion relationship between energy consumption and driving distance. Meteorological information includes temperature, relative humidity, precipitation, wind speed, wind direction, solar irradiance, and air pressure, and is spatially mapped or matched to the nearest location based on vehicle or station location to ensure that its time index is consistent with vehicle operation data; Traffic flow refers to a sequence of traffic states constructed based on road location and time index. Each time step in the traffic state sequence includes the number of vehicles passing through the road, the average speed, and traffic event markers.

3. The method for predicting the travel demand of electric official vehicles according to claim 1, characterized in that, By matching all time series variables in the first dataset with the mileage and power consumption sequences in the target sequence according to a unified time index, each variable in mileage, travel time, charging records, battery capacity, unit energy efficiency, meteorological information and traffic flow is used as a feature variable to form a feature correlation calculation set containing a one-to-one correspondence between each feature variable and the target sequence. The power consumption sequence includes the actual power consumption of the vehicle's charging records at each time step, arranged with a unified time index. The value is determined by subtracting the feedback power consumption in the same time period from the increase in the charging power in the vehicle's charging records at each time step, and then comparing and verifying the driving mileage with the unit energy efficiency. By calculating the Pearson correlation coefficient and partial correlation coefficient for each feature variable in the feature correlation calculation set and the target sequence respectively, correlation calculations are performed under linear and control conditions. The Pearson correlation coefficient is calculated by calculating the mean, standard deviation, and covariance of the feature variables and the target sequence at all time steps, and using the ratio of the product of the covariance and the standard deviation as the correlation coefficient. The calculated correlation coefficient is compared with a preset lower threshold, and feature variables with correlation coefficients not lower than the preset lower threshold are selected to generate an initial feature set.

4. The method for predicting the travel demand of electric official vehicles according to claim 3, characterized in that, Variance inflation is calculated for each feature variable in the initial feature set. A linear regression equation is established by using this feature variable as the dependent variable and the remaining feature variables as independent variables, and the coefficient of determination is calculated. and according to the formula Calculate the variance inflation value; Features identified as collinear variables are removed from the initial feature set, while features with variance inflation values ​​below a threshold are retained. The retained feature variables are re-aligned and standardized according to the time index to form a second feature set.

5. The method for predicting the travel demand of electric official vehicles according to claim 1, characterized in that, By arranging each time series in the second feature set according to a uniform time index to form a continuous input signal sequence, and performing local extremum identification at each time step, the distribution boundaries of the upper and lower envelopes of the signal are determined. The upper and lower envelopes are constructed as continuous functions through spline interpolation, and their mean functions at each time step are calculated. The local fluctuation signal is obtained by subtracting the mean function from the original input signal, and the extreme value detection and mean calculation are repeated until the change of the mean function of the local fluctuation signal is lower than the preset convergence threshold, thereby solving the first eigenmode function. After separating the first intrinsic mode function from the input signal, its residual part is retained as a new input signal, and the sieving and iteration process continues until the residual signal no longer contains local extrema, thereby solving the entire sequence of intrinsic mode functions and the final residual sequence in sequence.

6. The method for predicting the travel demand of electric official vehicles according to claim 5, characterized in that, For each intrinsic mode function, calculate its energy percentage and sample entropy. The energy percentage is calculated as the ratio of the sum of squares of the mode function to the sum of squares of all mode functions. The sample entropy is calculated as the average negative logarithm of the self-similar probability logarithms based on time delay embedding. Eigenmode functions whose energy percentage is between the lower and upper limits of a preset interval and whose sample entropy is between the lower and upper limits of a preset steady-state interval are marked as reserved modes. The eigenmode functions marked as preserved modes are weighted and summed using their energy percentages as weighting coefficients to generate a third time series matrix, which is then output.

7. A predictive system for the travel demand of electric official vehicles, characterized in that, include: The data construction module collects data on the mileage, travel time, charging records, battery capacity, unit energy efficiency, weather information, and traffic flow of electric official vehicles to obtain raw data aligned with a unified timestamp; and forms the first dataset based on the raw data. The feature selection module performs correlation calculations on the mileage and power consumption sequences in the first dataset and the target sequence to identify feature variables with correlation coefficients not lower than a preset lower threshold; it then performs variance inflation calculations on the identified feature variables and removes collinear variables according to a preset threshold to form a second feature set. The sequence decomposition and reconstruction module performs empirical mode decomposition on the time series in the second feature set to solve for the intrinsic mode function sequence and the residual sequence; it performs energy proportion statistics and sample entropy statistics on each intrinsic mode function, and selects the retained intrinsic mode functions based on the lower and upper limits of the energy proportion interval and the lower and upper limits of the sample entropy interval; it then sums the retained intrinsic mode functions according to their weights to generate the third time series matrix. The parameter optimization module constructs a parameter vector containing the number of hidden units in the Long Short-Term Memory network, the number of network layers, the sliding window length, the learning rate, and the target weight coefficients. Using an arithmetic optimization algorithm, with the globally optimal parameter vector in the current population as a reference, it performs global exploration using multiplication and division operations, and local development using addition and subtraction operations, updating the positions of candidate parameter vectors. The module calculates fitness using the root mean square error on the validation set as an evaluation metric, iteratively filters and solves for the optimal parameter vector, and outputs the optimized parameter set. The prediction generation module inputs the third time series matrix and the optimized parameter set into the long short-term memory network, performs forward computation and training updates, and solves the mileage sequence and power consumption sequence within a predetermined future time period; the solved prediction sequence is output according to the time index to generate the prediction result of electric official vehicle travel demand.

Images